Simulating how environmental interventions may change the effect of wildfire on acute-care utilization
A shiny app
Michael D Garber
2025-05-22
Link to shiny app: https://michaeldgarber.shinyapps.io/wf-emm/
1 Introduction
1.1 Background: Do et al. (GeoHealth, 2024)
Do and colleagues have recently estimated spatially varying effects of short-term exposure to wildfire PM2.5 on acute-care hospitalizations (Do et al. 2024) in California. They defined the exposure as a day with wildfire smoke PM2.5 ≥15 μg/m3. The outcome was the number of respiratory emergency department visits and unplanned hospitalizations.
They used a case-crossover design to estimate effects at the level of the zip-code tabulation area (ZCTA) throughout 1,396 ZCTAs in California from 2006-2019. A map of these spatially varying effect estimates, expressed as a rate difference per 100,000 person-days, appears below (adapted from their figure 5).
A positive rate difference (shades of green) indicates that wildfire had an estimated harmful effect (i.e., more) hospitalizations, while a negative value (purple shades) implies that wildfire led to fewer hospitalizations.
In addition, they explored community characteristics possibly driving this spatial heterogeneity. The authors state:
We used meta‐regression to evaluate potential effect modification by community characteristics on the effect of a wildfire smoke day on acute care utilization at the ZCTA level. For each community characteristic, which was selected a priori, we ran a meta‐regression of the pooled ZCTA‐specific rate difference on the community characteristic. To preserve statistical power, we excluded 100 ZCTAs without complete data for 14 community characteristics other than A/C prevalence, and we excluded 274 ZCTAs for meta‐regression of the A/C prevalence. Our estimates are reported as rate difference per interquartile range increase of the community characteristic.
Below, adapted from their figure 6, is a plot of the increase in rate difference per interquartile range (IQR) increase for each socio-demographic community characteristic they considered:
This analysis of effect modification describes how the estimated effect of wildfire smoke on acute-care utilization varies based on variation in the community characteristics. Results show, for example, that the wildfire-utilization effect was stronger in communities with lower levels of air conditioning and with higher levels of uninsured populations.
Interpreting a specific value from the figure, an increase in A/C proportion from the 25th percentile to the 75th percentile was associated with a 0.239 (95% confidence interval: 0.0672, 0.411) lower rate rate difference in the effect of wildfire smoke on same-day respiratory acute-care utilization.
1.2 Distinction between effect modification vs interaction
It is important to note that just because wildfire’s effects on acute-care utilization vary across these values, it does not necessarilly mean that intervenining to change these variables would necessarily change the wildfire-healthcare-utilization effect.
Tyler VanderWeele describes the important distinction between effect modification and interaction on p. 268 of his book, Explanation in Causal Inference (bold added for emphasis here):
If we found that the effect of our primary exposure varied by strata defined by the secondary factor in this way, then we might call this “effect heterogeneity” or “effect modification.” This might be useful, for example, in decisions about which subpopulations to target in order to maximize the effect of interventions. Provided that we have controlled for confounding of relationship between the primary exposure and the outcome, these estimates of effect modification or effect heterogeneity could be useful even if we have not controlled for confounding of the relationship between the secondary factor and the outcome. (Tyler J. VanderWeele 2015)
What we would not know, however, is whether the effect heterogeneity were due to the secondary factor itself, or something else associated with it. If we have not controlled for confounding for the secondary factor, the secondary factor itself may simply be serving as a proxy for something that is causally relevant for the outcome (Tyler J. VanderWeele and Robins 2007).
VanderWeele continues (bold again added for its relevance here):
If we are interested principally in assessing the effect of the primary exposure within subgroups defined by a secondary factor then simply controlling for confounding for the relationship between the primary exposure and the outcome is sufficient. However, if we want to intervene on the secondary factor in order to change the effect of the primary exposure then we need to control for confounding of the relationships of both factors with the outcome. When we control for confounding for both factors we might refer to this as “causal interaction” in distinction from mere “effect heterogeneity” mentioned above (T. J. VanderWeele 2009).
That is, effect modification can be considered a description of variation in an effect of a primary exposure over values over a secondary variable, whereas assessment of interaction requires the secondary factor to be considered as a causal factor.
To inform policy, it would be useful assess interaction to know what environmental factors could be intervened upon to mitigate wildfire’s health effects. Both tree canopy and air conditioning could plausibly attenuate wildfire’s health effects and in that way, interact with wildfire on the outcome. In the case of tree canopy, previous research has shown that green space can reduce air pollution.(Diener and Mudu 2021) And previous research in California has suggested that air conditioning use could mitigate the health effects of wildfires.(Stowell et al. 2025)
In addition, they are socio-environmental pathways that are relatively amenable to intervention. It would be a realistic policy option to encourage more tree planting, for example, or institute policies to encourage more air conditoning use during wildfire events.
However, as both of these factors may be associated with other factors on the causal pathway from wildfire to health, it would be important to control for potential confounders.
1.3 The need for locally developed health-impact assessment
Furthermore, many health-impact assessment studies make a strong transportability assumption–using dose-response functions developed elsewhere and applying it to the local context. There is a need for locally tailored health-impact assessment studies.
1.4 Analysis goals
The primary goal of this analysis is to assess how the distribution of the effect of wildfire onhealthcare utilization effect could change were there interventions changing environmental factors that influence the effect. We do so estimating the effect of tree canopy and air conditioning on the spatially varying rate difference.
A secondary goal, related to training and skill development, is to explore the utility of R Shiny tools for presenting high-dimensional results with which a user can interact. For example, a user could choose to raise the distribution of the effect modifier to the 50th percentile and see how results might change.)
2 Methods
The overall approach is as follows:
- Estimate the effect of both tree canopy and air conditioning (separately) on the spatially varying rate difference, controlling for potential confounders that may influence these associations such as urban vs rural status, biome, and socioeconomic characteristics. In these models, we will also consider interaction variables, allowing the relationship between these main exposures (tree canopy and AC) and the rate difference to vary across strata of a third variable.
- Use these models to simulate the overall burden and spatial distribution of healthcare utilization attributable to wildfire that could be averted by intervening upon these variables.that
2.1 Data exploration
To inform modeling, we explored the distribution of all variables as well as bivariate associations between tree canopy, air conditioning, the rate difference, and other covariates.
2.1.1 Univariate distributions
2.1.1.1 Outcome: Rate difference per 100,000 person-days
## # A tibble: 4 × 3
## ruca_cat rd_100k_pt_med rd_100k_pt_mean_wt
## <fct> <dbl> <dbl>
## 1 (0,3] -0.0331 0.196
## 2 (3,6] -0.0351 -0.164
## 3 (6,9] -0.522 0.141
## 4 (9,10] -0.732 -0.306
2.1.1.2 Intervention variables: Tree canopy and air conditioning
2.1.1.3 Proportion above poverty and proportion insured
2.1.1.4 Biome and rural-urban classification
Definitions of rural-urban commuting codes:
- Metropolitan area core: primary flow within an urbanized area (UA)
- Metropolitan area high commuting: primary flow 30% or more to a UA
- Metropolitan area low commuting: primary flow 10% to 30% to a UA
- Micropolitan area core: primary flow within an Urban Cluster of 10,000 to 49,999 (large UC)
- Micropolitan high commuting: primary flow 30% or more to a large UC
- Micropolitan low commuting: primary flow 10% to 30% to a large UC
- Small town core: primary flow within an Urban Cluster of 2,500 to 9,999 (small UC)
- Small town high commuting: primary flow 30% or more to a small UC
- Small town low commuting: primary flow 10% to 30% to a small UC
- Rural areas: primary flow to a tract outside a UA or UC